[linux-sunxi] Re: [PATCH v5 4/4] crypto: Add Allwinner Security System crypto accelerator

[clabbe.montjoie@gmail.com (Corentin LABBE)]

Le 22/10/2014 11:00, Arnd Bergmann a ?crit :
> On Sunday 19 October 2014 16:16:22 LABBE Corentin wrote:
>> Add support for the Security System included in Allwinner SoC A20.
>> The Security System is a hardware cryptographic accelerator that support AES/MD5/SHA1/DES/3DES/PRNG algorithms.
>>
>> Signed-off-by: LABBE Corentin <clabbe.montjoie@gmail.com>
> 
> Please wrap lines in the changelog after about 70 characters.
> 

Oups I just see the corresponding part in submittingpatches.txt
Sorry

>> --- /dev/null
>> +++ b/drivers/crypto/sunxi-ss/sunxi-ss-cipher.c
>> @@ -0,0 +1,489 @@
> 
>> +#include "sunxi-ss.h"
>> +
>> +extern struct sunxi_ss_ctx *ss;
> 
> 'extern' declarations belong into header files, not .c files. It would
> be even better to avoid this completely and carry the pointer to the
> context in an object that gets passed around. In general we want drivers
> to be written in a way that allows having multiple instances of the
> device, which the global pointer prevents.
> 

As I already said I think the driver will never be used with multiple instance.
But since many people want this pointer dead, I will work on it.

>> +
>> +	src32 = (u32 *)src_addr;
>> +	dst32 = (u32 *)dst_addr;
> 
> 
> You appear to be missing '__iomem' annotations for the mmio pointers.
> Please always run your code through the 'sparse' checker using 'make C=1'
> to catch and fix this and other erros.
> 

Ok, but with which version of sparse do you have such a warning. I use the 0.5.0 version and I got no warning at all.

>> +	ileft = areq->nbytes / 4;
>> +	oleft = areq->nbytes / 4;
>> +	i = 0;
>> +	do {
>> +		if (ileft > 0 && rx_cnt > 0) {
>> +			todo = min(rx_cnt, ileft);
>> +			ileft -= todo;
>> +			do {
>> +				writel_relaxed(*src32++,
>> +						ss->base +
>> +						SS_RXFIFO);
>> +				todo--;
>> +			} while (todo > 0);
>> +		}
> 
> This looks like it should be using writesl() instead of the 
> writel_relaxed() loop. That should not only be faster but it will
> also change the byte ordering if you are running a big-endian
> kernel.
> 
> Since this is a FIFO register, the ordering that writesl uses
> is likely the correct one.

Great, the code is much cleaner with it. (with up to 10% speed gain)

Thanks

Corentin